Search for Bs decays to tau lepton pairs with the CMS experiment at the CERN LHC

The high energy physics experiments at the LHC are designed to address many fundamental questions in modern physics. Extracting the relevant information from the collected data that can answer these questions is a difficult challenge due to the complexity and the high dimensionality. The emergence of deep learning algorithms have advanced the state of the data analysis methods by enabling the extraction of higher-level features and consequently reducing the dimensionality, which is a crucial improvement considering the vast size of collision data that is necessary to observe rare physics processes of interest. Within the scope of this thesis several machine learning techniques have been implemented to study the rare Bs → ττ decay into tau leptons with the two tau decay modes τ → ντµνµ and τ → πππντ respectively. To this purpose, B-Parking data containing a large number of Bs mesons, acquired by CMS during the Run 2 of the LHC and simulated Monte Carlo samples that include the decay channel of interest, have been used. The reconstructed events are filtered for the specific decay signature by a graph neural network that classifies triplets of charged particles as candidates for the 3-prong tau decay τ → πππντ for events that are triggered by a muon, which is the candidate muon for the τ → ντµνµ decay. Identifying this decay channel is complicated by the escape of at least three neutrinos; two of which are produced in the 3-prong decay and the third in the semi-hadronic decay. Neural networks and gradient boosted decision trees have been explored as methodologies to recover the lost information from the measured momenta of the visible particles. Two supervised learning methods have been implemented; regressions to the four-momentum of the semi-hadronic and 3-prong decaying tau with the goal of estimating the four-momentum of the originating Bs meson and a classification between the signal and background events. Furthermore, a semi-supervised learning algorithm has been designed to complement the supervised classifier.